The Oldbury projectA cache for 3,815 Sandwell streets, and what it gets wrong
An agent answers "which postcode district is this street in?" and remembers its answers. The question is what to remember them by.
The learner pulls every named road inside the Sandwell boundary from OpenStreetMap with one Overpass request and keeps 9,485 road segments. Each segment is given the postcode district of the nearest of Sandwell's 6,906 postcodes, typically about 37 m from its middle. Segments with the same name whose centres lie within 400 m of each other are joined into one street. That leaves 3,815 streets but only 3,450 different names, because 263 names are used by two or more separate streets. High Street appears eight times, in six districts; Park Street also spans six districts and Church Street four.
Now the agent. It receives 20,000 questions, each about a street picked at random, and must answer with a postcode district. Looking the answer up properly always works but costs effort, so the agent keeps a cache: before looking anything up it checks whether it has answered an "equivalent" question already, and if so it reuses that answer. The only thing that changes between our four agents is what counts as equivalent, in other words the cache key. We ran each agent 20 times with different random question streams.
Four cache keys, 20,000 questions per run, averaged over 20 runs (our Python)| Cache key | Questions answered from the cache | Answers that were wrong |
|---|
| Exact street name | 82.8% | 7.90% |
| Name without its road-type word ("Church Road" and "Church Lane" count as one) | 86.7% | 23.39% |
| First word only | 88.9% | 31.15% |
| Exact name plus the ward it is in | 81.2% | 0.11% |
Even the strictest-looking key, the exact name, gets almost one answer in twelve wrong, because a remembered High Street in DY4 is handed out for the High Street in B64. Loosening the key, which is what a "semantic" cache does, buys four to six more points of cache use and roughly triples or quadruples the error. The fix runs the other way. Adding the ward, a piece of context the answer actually depends on, costs 1.6 points of cache use and cuts wrong answers to about one in a thousand. The leftover errors come from the few streets whose ward and district do not line up, and the learner is asked to find them.
Ages 8 to 11
Play a memory game where the answer card depends on two clues, then play it again with one clue hidden.
Ages 11 to 15
Build a dictionary cache in Python for a list of streets and count its hits and mistakes.
Ages 15 and up
Load the real Sandwell data, test four cache keys over 20 runs, and explain every remaining error.
Sources and honesty notes
Road names © OpenStreetMap contributors (Open Database Licence). Postcode positions contain OS data © Crown copyright and database right 2026 and Royal Mail data © Royal Mail copyright and database right 2026. Joining segments into streets by a 400 m rule is our own simplification, so one long road with a gap can count as two streets. The agent, the questions and every percentage come from our simulation.
